Topics for Theses and Projects

Topics for Bachelor Theses, Master Theses, Projects and Lab Rotations

Below is a list of thesis and project ideas. Projects marked as Open are available for new students. If interested consult the instructions on the Thesis Writing Guidelines page and write me an Email. In the Email, attach your transcript of records and name a few topics you are interested in. Own project ideas are welcome too.

Open

  • Open: Learning binary decision diagrams for combinatorial optimization with RL.
  • Open: Improving TRM4CO (tiny recursive models for combinatorial optimization): reduce the dependence on problem-specific input encoders and output decoders and make the method generic across problems.
  • Open: Automated discovery: Apply evolutionary loss search to other applications — RL algorithms, optimizers, etc.
  • Open: Automated discovery: Improved evolutionary strategies — cross-over operators, ideas from evolutionary algorithm literature.
  • Open: Robust machine unlearning: Adversarial loop where an attacker finds jailbreaks and the model unlearns on them.
  • Open: Prompt optimization: PIAST in-context example synthesis + prompt rewriting + order optimization.
  • Open: Self-Distillation with teacher having few-shot prompts, analogous to ground truth solutions in SDPO or OPSD.
  • Open: Orthogonalization of MLPs in neural networks, see the Thinking Machines blog post Modular Manifolds.
  • Open: Memory in neural networks: Engrams and indexing of latent states.
  • Open: Mini League of Legends: Develop a fast, miniaturized version of League of Legends on GPU for fast training of RL agents.
  • Open: Reverse-engineering: Reimplement some state-of-the-art complex software, like simplex and other optimization algorithms.
  • Open: Hard verification tasks by construction: Automatically generate Lean verification tasks by chaining semantics-preserving optimizations of simple reference programs, plus planted bugs for prove-or-disprove, calibrated so that current LLMs struggle. Use them as an RL environment for training provers. Starting points: EquiBench, Learning to Disprove.
  • Open: Hard crafting worlds with guaranteed solutions: Procedurally generated text games with invented items, hidden recipes, consumable keys, tool durability and scarce resources. A planner or CP-SAT solver certifies solvability and the optimal plan length, and deliberate traps (actions that make the world unsolvable) make them hard. Use as RL environment for LLM agents. Related: AgentOdyssey, Learn2Play, TextCraft.

Ongoing

  • Ongoing, Doguhan Bahcivan: NanoRTS: Massively parallel reimplementation of MicroRTS on GPU in JAX.
  • Ongoing, Ogulcan Asik: Craftax with LLMs: LLM-generated skills for Craftax.
  • Ongoing, Hani Ouraghene: Vertex tracking: Graph neural networks and MaskFormer for clustering detections into vertices.
  • Ongoing, Torben Glass: Parallel coordinated Tiny recursive models: Scale test-time compute by sampling multiple reasoning traces and improve performance by exchanging information between them.
  • Ongoing, Julius Glamann: Looped transformers: Improve LLMs by searching efficiently for layer repetitions, building on Neuroanatomy.
  • Ongoing, Iraj Masoudian: SynKE: Knowledge editing in LLMs by training on synthetic data from self-play.
  • Ongoing, Edward Ittermann: Enhancing incident ticket management with unsupervised AI-driven categorization (with Bosch Building Technologies).
  • Ongoing, Julius Lilie: Modeling grid topology for redispatch forecasting with graph neural networks in the German energy market (with Uniper).
  • Ongoing, Esther Schünemann: Multi-state models for predicting treatment resistance in major depressive disorder (with Fraunhofer SCAI).

Finished

  • Finished, Devrim Kekec: Particle Imaging Velocimetry (PIV) with event-based cameras and multicut (RAMA).
  • Finished, Malik Mardan: Tiny recursive models for combinatorial optimization.
  • Finished, Moritz Ohrt: Prior fitted networks for finding data-efficient policies for RL.
  • Finished, Ria Choudhary: Rational basis functions in SplineCNN for secondary vertex reconstruction, improving MaskFormer for vertex tracking.
  • Finished, Jonas Friesen: Fast running robotic dog with RL locomotion policies in the Genesis simulator.
  • Finished, Daniel Dratschuk: Transcoda: End-to-end music sheet OCR with synthetic training data (NeurIPS 2026).
  • Finished, Oussama Al-Meziani: Reasoning capabilities of base LLMs on ARC-AGI.
  • Finished, Anthony Nascente: A systematic evaluation of retrieval strategies for domain-specific RAG (with Senacor).
  • Finished, Anthony Nascente: Relational deep learning for customer data in telecommunications.
  • Finished, Niklas Glaser: MicroRTS with league training: Improve the RL-trained initial MicroRTS agent with self-play and league training.
  • Finished, Malte Günnewig: Perturbing costs for solving ILPs with linear relaxation.
  • Finished, Paweł Batorski: EvoMU: Evolutionary search for machine unlearning losses.
  • Finished, Patryk Rybak: REBEL: Jailbreaking unlearned models.
  • Finished, Bartosz Dziuba, Kacper Kuchta: TATRA: Training-free instance-adaptive prompting through rephrasing and aggregation.
  • Finished, Paweł Batorski: PIAST: Fast in-context example synthesis for prompt engineering.
  • Finished, Paweł Batorski: GPS: Instance-specific unsupervised prompt engineering.
  • Finished, Paweł Batorski, Adrian Kosmala: PRL: Reinforcement learning for automatic prompt engineering.
  • Finished, Abtin Pourhadi: Normalized Matching Transformer: Keypoint matching for sparse semantic matching.
  • Finished, Antonio Millitello: Keypoint matching for multiple images with cycle consistency.
  • Finished, Noah Yildiz: Dense keypoint matching with contrastive learning.
  • Finished, Tom Sanchez Deutsch: 3D point cloud segmentation with state space models.
  • Finished, David Seiler: Kolmogorov-Arnold convolution networks.
  • Finished, Emre Gökcek: Walking policy with RL: Train a robotic dog to walk through difficult terrain and make artistic jumps in the Genesis simulator.
  • Finished, Maximilian Huft: MicroRTS with RL: Behavioral cloning and PPO against bots to play MicroRTS.